More Podcasts by InfoQ episodes

Educating an Architect: A Conversation with Luca Mezzalira thumbnail

Educating an Architect: A Conversation with Luca Mezzalira

Published 2 Nov 2026

Duration: 01:07:21

Architects must evolve from technical designers to social-technical enablers, using domain-driven design to balance AI automation with human oversight in core systems.

Episode Description

In this podcast Michael Stiefel spoke to Luca Mezzalira about software architecture as a socio-technical discipline that balances technological design...

Overview

The conversation explores how architecture emerged as a natural evolution of the engineering role rather than a formal training path, with the guest describing how refactoring, system design, and understanding business context gradually shaped his professional identity. He emphasizes that architecture is fundamentally a social-technical discipline, where the architect's primary function is balancing decisions across organizational communication structures rather than being the most technical person in the room.

The discussion turns to how AI is reshaping architectural practice, with the guest proposing a domain-driven design framework for determining how much human involvement each system domain requires. He advocates for decentralizing architectural decision-making through guardrails and coaching rather than centralized design, while acknowledging that cross-domain features still require someone to look holistically. The conversation also covers education through collaborative katas and case studies, feedback culture as a core competency, and the practical challenges of maintaining craft skills in an AI-assisted development environment.

What If

  • What if you mapped your solo project into core, supporting, and generic domains to set different AI delegation levels for each?

    • Move: Identify which parts of your system are core (your unique value), supporting (expected but not differentiating), and generic (commodity integrations like payments), then apply heavier human review to core and lighter review to generic.
    • Why Now? The guest describes this as his current mental model for deciding "how much human in the loop I want to have," noting that generic domains like payments involve "no much novelty" and can tolerate more AI delegation.
    • Expected Upside: You could propose a workflow where boilerplate integrations are handled more by AI while your core logic retains full human control, potentially reducing your time on repetitive work while protecting quality where it matters most.
  • What if you treated architecture friction in your system as a diagnostic signal rather than a problem to fix immediately?

    • Move: When your codebase resists change, pause and run a context analysis of your constraints (business direction, team structure, technical debt) before redesigning any component.
    • Why Now? The guest states "Friction, I always interpret as a gift because it's a feedback that architecture provides you. It's telling you something doesn't work. And that something might not be in the architecture space."
    • Expected Upside: You might discover the root cause is organizational or contextual rather than technical, avoiding unnecessary rewrites and preserving working architecture that is actually signaling a different problem.
  • What if you built a structured two-pass review process for your own AI-generated code to maintain craft skills?

    • Move: First pass: verify correctness and test coverage of AI output. Second pass: evaluate it against your system's architectural invariants, asking whether abstractions are leaking or whether the code would survive a future refactor.
    • Why Now? The guest notes that "two people should review a PR" and that this is "even more important now" with AI-generated code, and that "the mastering part is trying, making an error."
    • Expected Upside: You could develop the intuition and "battle scars" the guest says are essential for senior judgment, while still benefiting from AI's speed, creating a sustainable learning loop that does not depend on a team of reviewers.

Takeaway

  • Map your system into core, supporting, and generic domains before delegating any work to AI, applying full human control to core logic and lighter review to commodity integrations like payment SDKs.

  • When your architecture resists change, diagnose the organizational and business context before redesigning; the guest interprets friction as "a feedback that architecture provides you" that may point to a non-technical root cause.

  • Decouple feedback from its delivery form: treat every comment, regardless of tone, as "a step closer to improving your solution" and use it to refine your design rather than reacting defensively.

  • Schedule regular collaborative design sessions (pair programming, structured code reviews, or self-directed design critiques) to maintain craft skills alongside AI-assisted development, since "the mastering part is trying, making an error."

  • Before adopting any architectural pattern or best practice, verify it fits your specific context; the guest warns that "not necessarily embrace every single best practice because it might not fit inside their context."

Recent Episodes of Podcasts by InfoQ

More Podcasts by InfoQ episodes